Group Behaviour Analysis of London Foot Patrol Police
Jianan Shen, Tao Cheng
SpaceTimeLab for Big Data Analytics, Department of Civil, Environmental and Geomatic
Engineering, University College London,
Gower Street, London WC1E 6 BT, UK
Nov 03, 2014
Summary
The main objective of this research is to propose a method for group movement pattern generalisation
and classification. To this end, DBSCAN is used on stay points for POI identification. Then,
movement features are extracted and selected for the behaviour classification. A kernel-based Support
Vector Machine method is developed to infer the working types of the officers based on the selected
features depicting their movement histories. By analysing the geo-tagged police data, we demonstrate
how this method can be used to reveal user information, especially interest information based on their
POIs and spatial-temporal movement patterns.
KEYWORDS: Policing, Travel Pattern, Machine Learning, SVM Classification, POI
1. Introduction
The more and more ubiquitously used GPS-integrated devices have generated increasingly large
movement history data than ever before. Such movement history data have enabled researchers to
analyse and visualise movement trajectories, home range and POI (Points of Interests) distribution,
and mobility patterns. Police foot patrol activity, as a special kind of human movement phenomenon,
can also be analysed in similar methods with certain modifications.
In this research, we intend to present a framework of patrol pattern analysis and classification of
police working type, which is capable of 1) extracting police POIs and traveling sequences from GPS
trajectory data; 2) selecting and summarising individual and group movement features for description
and comparison of officers undertaking different works; 3) SVM classification of different types of
officers based on the selected moving history features.
2. Case Study and Data
The case study takes place in the Camden Borough of London(Figure 1). Five major police stations
are located (Places No. 1, 7, 8, 9 and 11 in Figure 1) in this region, namely, West Hampstead,
Hampstead, Kentish Town, Albany Street and Holborn.
Automatic Personnel Location Systems (APLS) data set provided by the Metropolitan Police
recorded officers’ location stamps collected by GPS-integrated radio sets portable on every officer in
patrol. The length of the data is 84 days. The 241525 records in the dataset recorded information such
as call signs, device IDs, as well as the locations and times of all 745 officers active in that period.
Usually, the sampling rate of the system is one update every 10 minutes.
Figure 1 The POIs identified in Camden, including 10 non-police-station POIs (1 of them locates
outside Camden) and 5 police stations.
3. Methodology and Results
3.1. POIs and Traveling Sequences Extraction
In discovering POIs, the location and time of officers' stopping behaviour is of more interests than the
moving factors (Palma, 2008). To this end, DBSCAN is used for the clustering of stay points where
officers stop or move slowly for more than 20 minutes. The massive activities and stops observed
nearby police stations are common reflection of police daily routine and office works, and hence
disturbed the searching for the POIs during the patrol activities. Therefore, outliners and records
generated within 400m radius from the police stations are removed before the density based
clustering. Only trajectory moving out of the radius will be considered as a start of a journey and
when the time interval between two records exceed 2 hours, the series will be separated as two trips.
By adjusting the parameters according to the quality of the data set and the method suggested by
Zhou (2012), 10 non-police-station POIs are discovered. For instance, POI No. 2 represents a tube
station, POI No.3 is considered to be a major road intersection.
Figure 2 Individual movement history expressed by sequences of visited POIs
By identifying building information on Ordinance Survey Maps and communicating with the police.
The main topics of these POIs are identified as schools, places of worships, tube stations and major
road intersections. The movement histories of the officers can then be simplified and expressed by a
sequence of common POIs the officers visited and the time they arrive and leave each POI attached in
the sequence (Figure 2). Similarly, it can also be expressed as a series of topics an officer went
through during his/her patrol journey. For example, the journey of Officer 2 in Figure 2 can be
summarised as "Commercial Area> School > Fire Station" and the attached time information about
how long s/he stayed in each POI. The information of time series, POI topics and trajectories will be
further processed to support the classification works.
3.2 Individual and Group Movement Features
After acquiring the semantic meaning of discovered POIs in previous step, individual and group
movement can be depicted and summarised with 14 features that are selected or extracted from the
raw GPS log as well as the summarised POI visiting history. The features include: mean and standard
deviation of speed, mean and standard deviation of patrol covering area, proportion of stopping time,
proportion of time spent in police stations and POIs of different topic types, proportion of time spent
in weekends and nights and the activeness expressed by the number of records generated by the
officer in unit time and so on.
Officer Group A
Officer Group C
Figure 3 The comparison of POI visiting frequency between two groups consisting of two different
types of officers in one month
Figure 4 shows 10 example features of the 3 groups. The values of features are z-score normalised for
further classification works. It can seen in the chart that the selected features of the 3 groups vary
greatly. For example Group B has strong daytime activeness proportion and seldom work at night
while the standard deviation of Group A is much larger than other groups.
Activeness
1
Time in
commercial… 0.5
Time on
major roads
Area covered
SD of speed
0
-0.5
-1
Average trip
distance
Daytime
ratio
In-station
ratio
Weekday
ratio
Stop-time
ratio
Group A
Group B
Group C
Figure 4 Radar chart comparison of 10 example features depicting the patrol behaviour of different
groups
3.3. SVM Classification of Officer works using RBF Kernel
Many researches proposed SVM and other machine learning methods analysing online user behaviors.
Some researchers developed similar methods on the analysis of travelling behaviour in real life, such
as Baraglia (2013).
By using the parameter selection method developed by Schölkopf (2002), RBF kernel is used in the
SVM classification of officer working types based on the individual movement features extracted in
previous section. The model is trained with officer movement data of which the working types have
already been labeled and is used to classify the data generated in 10 days of the next month. The
accuracies of this attempt are logged in Table 1.
Table 1 Classification accuracy by training sets of different sizes
Size of training sets
10 days
20 days
30 days
Training Error
13.52%
15.00%
14.83%
Accuracy of classification
68.89%
72.34%
73.42%
This classification is based on manually-selected and unoptimised feature sets. This method will be
further improved with optimum feature selection and using more advanced kernels.
4. Summary and future work
In this research, a method for identifying user (officer) type based on travel (patrol) history is
proposed. The framework include density-based POI clustering, feature selection and validation, as
well as machine learning classification. The Camden APLS data enabled the study that others cannot
proceed with due to the lack of modern GPS-enabled policing equipments. The method revealed the
movement features of different officers in space and in time. It can also be used in other time series
geo-tagged data for automatic movement pattern generalisation, traveler interest and routine mining,
similarity analysis and so on.
Further works may include improving the performance of SVM in unbalanced data sets (the officer
numbers in different types vary greatly) and comparing SVM with other methods such as KNN and
Artificial Neural Networks.
5. Acknowledgements
This work is part of the project - Crime, Policing and Citizenship (CPC): Space-Time
Interactions of Dynamic Networks (www.ucl.ac.uk/cpc), supported by the UK Engineering and
Physical Sciences Research Council (EP/J004197/1). The data provided by Metropolitan Police
Service (London) is highly appreciated.
6. Biography
Jianan Shen is a PhD student in UCL SpaceTimeLab. He studied Information Engineering in NUDT
in China (BEng, 2013). He is now working on the Crime, Policing and Citizenship Project sponsored
by the EPSRC, UK. His research interests include movement pattern and trajectory analysis with
Machine Learning approaches.
Tao Cheng is a Professor in GeoInformatics, and Director of SpceTimeLab for Big Data Analytics
(http://www.ucl.ac.uk/spacetimelab), at University College London. Her research interests span
network complexity, Geocomputation, integrated spatio-temporal analytics and big data mining
(modelling, prediction, clustering, visualisation and simulation), with applications in transport, crime,
health, social media, and environmental monitoring.
References
Baraglia R, Muntean C I, Nardini F M and Silvestri F (2013). LearNext: learning to predict tourists
movements. In Proceedings of CIKM, 751-756.
Palma A (2008). A Clustering based Approach for Discovering Interesting Places in Trajectories.
Symposium on Applied Computing, March 16-20, Fortaleza, Brazil. 863-868.
Schölkopf B and Smola A J (2002). Learning with Kernels: Support Vector Machines, Regularization,
Optimization, and Beyond. MIT Press.
Zhou C (2004). Discovering Personal Gazetteers: An Interactive Clustering Approach. Geographic
Information Science, 266-273.